cairn: Skill for Claude Code

Reconcile the remaining plan

Loaded by cairn-dev loop mode after Verify and before Land. Inputs: the selected unit (slug or finding code), its resolved node, its validated todo body when it has one, the proof Verify produced, and the bound $CAIRN.

A campaign runs across many fresh sessions with no memory between them. The plan the next session reads is whatever this commit leaves on main. Much of the work is research based, so a landed unit routinely invalidates a later one. Recording that is part of the unit, not paperwork after it.

Declared exits, exactly one, as the last line you return to loop mode:

• RECONCILED: the plan on disk matches this unit's evidence, and the edits sit in the worktree for Land to stage.
• LOOP HALTED: reconciling needs a maintainer ruling this session cannot make.

Provenance only. This recipe never selects the next unit, never repeats, never retries, never interprets a terminal token, and never commits or pushes. Land owns the commit; the invoking user or harness owns repetition (dec.no-orchestrator, dec.loop-command-harness-model).

Authoring paths

Every obligation below has one path. $CAIRN research, $CAIRN todos, and $CAIRN decisions are read-only queries, not creation verbs.

Obligation	Path
Observation and result	hand-author meta/research/<slug>.md (id res.<slug>, nodes: required)
A rule changed	$CAIRN decision new <slug> --node <id>, then author the body
Status change	$CAIRN todo set <slug> <open|in_progress|done|blocked>
New prerequisite	$CAIRN todo new <slug> --node <id>
Todo body, Depends on, umbrella Waves, change spec	direct file edit

1. Record the observation

If this unit produced evidence (a measurement, a rejected option, a behaviour the plan did not predict), write it as a Research artefact naming the node it constrains. If it changed a rule rather than only informing one, write a Decision as well and state retain, revise, or remove for the hypothesis the unit tested.